Intelligence as we know it isn't a unitary thing.
You don't have one mind. Your brain is full of non-conscious features. Impulses from your optic nerve are fed into the visual part of your brain and are heavily processed before anything resembling your consciousness gets access to them.
Your visual experience is a hallucination; most of your vision is basically in black and white, the only part with any resolution is directly where you are looking, you have blind spots, your nose is (typically) filtered out, your vision turns off for a fraction of a second every time your eye moves, etc. Lines, angles, edges, circles; pattern recognized and hallucinated.
This "low level" visual processing occurs without bothering your "mind", and it happens in your brain.
If we build our AIs based off of the only intelligence we know, we'll probably solve various subproblems and connect them up. The visual processing system won't be a general AI; it will be a specialized AI that does high quality visual processing.
The "feed" to the "really smart" part of the AI's brain is thus not going to look like what the camera sees. It is going to be marked up with information that the specialized visual processing unit has worked out.
Possibly multiple layers of such specialized visual processing units are going to be turned into a pipeline. Some of them will recognize humans and highlight them. This means that the "mind" doesn't have to have the ability to rapidly and reliably notice humans in its field of view; it just has to have good enough visual processing to recognize the highlights.
A HUD could exist for similar reasons. The intelligent part of the AI is human-like; it no more aware that it processes binary data than our brain is aware that it is processing glucose or action potentials. "Lifting" data up to the level that the AI experiences (visual bullet count, etc) instead of providing a binary feed could make it easier for the "command AI" to pay attention to it.
A side benefit to all of this is that the augmentations would work with a human for both testing and development. Other benefits include the fact that the visual coprocessing units can also be modularly replaced without rebuilding the AI In addition, and if the visual co-processing/HUD systems are broken the core AI has limited visual processing and can proceed without them.